Ray-casting implementations require that the connectivity between the cells of the dataset to be explicitly computed and kept in memory. This constitutes a huge obstacle for obtaining real-time rendering for very large models. In this paper, we address this problem by introducing a new implementation of the ray-casting algorithm for irregular datasets. Our implementation optimizes the memory usage of past implementations by exploring ray coherence. The idea is to keep in main memory the information of the faces traversed by the ray cast through every pixel under the projection of a visible face. Our results show that exploring pixel coherence reduces considerably the memory usage, while keeping the performance of our algorithm competitive with the fastest previous ones.

1. Memory-Aware and Efficient Ray-Casting Algorithm
Saulo Ribeiro1 , Andr´ Maximo1, Cristiana Bentes2 , Antˆ nio Oliveira1 and Ricardo Farias1
e o
1 2
COPPE - Systems Engineering Program Department of Systems Engineering
Federal University of Rio de Janeiro - Brazil State University of Rio de Janeiro - Brazil
Cidade Universit´ ria - CT - Bloco H - 21941-972
a Rua S˜ o Francisco Xavier, 524 - 20550-900
a
{saulo, andre, oliveira, rfarias}@lcg.ufrj.br cris@eng.uerj.br
Abstract the traveling of a ray through out the mesh is guided by the
connectivity of the cells of the mesh, avoiding the need of
Ray-casting implementations require that the connectiv- sorting the cells. The disadvantage is that cell connectivity
ity between the cells of the dataset to be explicitly com- has to be explicitly computed and kept in memory. In other
puted and kept in memory. This constitutes a huge obsta- words, the amount of memory used by ray-casting meth-
cle for obtaining real-time rendering for very large mod- ods is a huge obstacle for handling very large models.
els. In this paper, we address this problem by introducing Recent previous efforts towards improving ray-casting
a new implementation of the ray-casting algorithm for ir- performance have focused on reducing the execution time
regular datasets. Our implementation optimizes the mem- by exploiting hardware graphics, or GPUs (Graphic Pro-
ory usage of past implementations by exploring ray coher- cessing Unit). This approach, however, only provides in-
ence. The idea is to keep in main memory the information of teractive rendering time when the whole dataset fits into
the faces traversed by the ray cast through every pixel un- the GPU memory. Once the data size exceeds the onboard
der the projection of a visible face. Our results show that memory, expensive transfers from main memory to the GPU
exploring pixel coherence reduces considerably the mem- have great impact on the rendering speed.
ory usage, while keeping the performance of our algorithm Therefore, the problem of rendering large datasets (that
competitive with the fastest previous ones. do not fit in GPU memory) must be addressed by soft-
ware solutions, due to the inherent larger memory capac-
ity [6, 11, 7, 12]. Only a few software solutions, however,
deal with irregular meshes. Garrity [5] proposed the first
1. Introduction method for ray-casting irregular meshes using the connec-
tivity of cells. Bunyk et al. [2] later improved Garrity’s
Direct volume rendering has become a popular technique work providing a faster algorithm. Pina et al. [13] improved
for visualizing volumetric data from sources such as scien- Bunyk approach in: memory consumption; completely han-
tific simulations, analytic functions, and medical scanners, dling degenerate cases; and dealing with both tetrahedral
such as MRI, CT, among others. The main advantage of di- and/or hexahedral meshes. By using new data structures,
rect volume rendering is to allow the investigation of the in- Pina et al. reduced significantly the memory requirements
terior of the data volume, exhibiting its inner structures. of the Bunyk approach, but the algorithms proposed were
Several algorithms and methods have been proposed for still slower than Bunyk.
direct volume rendering, the most popular one is the ray- In this work, we propose a novel ray-casting algorithm
casting algorithm. In this algorithm, a ray is cast through based on Pina et al. approaches, reducing even more the
each pixel of the image from the viewpoint. The trace of memory consumption and improving the execution time.
the ray determines which cells of the volume each ray inter- Our approach is based on the fact that the decision of how
sects. Every pair of intersections is used to compute the cell to store the face information is the key for memory con-
contribution for the pixel color and opacity. The ray stops sumption and execution time. The more information is kept
when it reaches full opacity or when it leaves the volume. in memory, the faster the algorithm computes the next in-
Compared to other direct volume rendering methods, the tersection of the ray. In this work, we tackle this tradeoff
great advantages of ray-casting methods are: the computa- by exploring ray coherence in order to maintain in mem-
tion for each pixel is independent of all other pixels; and ory only the face information of the most recent traversals.

2. Our idea is to use each visible face computed in the pre- entry point by rendering front faces, and then traverse
processing step to guide the creation and destruction of in- through cells using the fragment program by storing the
ternal faces data in memory. Our algorithm, called Visible cells and connectivity graph in textures. Their method, how-
Faces Ray-casting (VF-Ray), obtained consistent and sig- ever, work only on convex unstructured data. Bernardon
nificant gains in memory usage over previous approaches. et al. [1] improve Weiler’s approach, by using Bunyk’s
We also reduced the execution time of Pina et al. approaches algorithm as the basis for the hardware implementation
by making a better use of the cache. When compared to the and depth peeling to correctly render non-convex irregu-
fast and memory expensive approach by Bunyk et al., we lar meshes. Espinha and Celes [3] propose an improvement
obtained comparable performance, spending only from 1/3 in Weiler’s work. They use partial pre-integration and pro-
to 1/6 of the memory. vide interactive modifications of the transfer function. They
The remainder of this paper is organized as follows. In also use an alternative data structure, but as their work is
the next section we discuss direct volume rendering of ir- based on Weiler’s work, the memory consumption is still
regular meshes. Section 3 briefly describes the Bunyk’s and high. Recently, Weiler el al. [17] proposed an improvement
Pina’s approaches. Section 4 describes our ray-casting al- over their previous work, where they deal with the prob-
gorithm and the improvements we made over previous ap- lem of storing the whole dataset in GPU memory by using
proaches. In section 5, we present the results of our most a compressed form, tetrahedral strips.
important experiments. Finally, in section 7, we present our In terms of other direct volume rendering algorithms dif-
conclusions and proposals for future work. ferent from ray-casting, it is important to mention the cell
projection approach. In this approach each polyhedral cell
of the volumetric data is projected onto the screen, avoid-
2. Related Work ing the need of maintaining the data connectivity into the
memory, but requiring the cells to be first sorted in visibil-
Within the research area of accelerating ray-casting
ity ordering and then composed to generate their color and
methods, two main research streams have emerged: soft-
opacity in the final image. In this scope, there are some soft-
ware approaches and graphics hardware approaches.
ware implementations, e.g., [10] and [4], that provide flex-
The very first implementation of the ray-casting algo-
ibility and easy parallelization. The GPU implementations
rithm for rendering irregular meshes was proposed in
of projection algorithms, however, are more popular [16, 9].
software by Garrity in [5]. His approach used the connec-
tivity of cells to compute the entry and exit of each ray. This
work was further improved by Bunyk et al. [2], by com- 3. Previous Approaches
puting for each pixel a list of intersections on visible
faces, and easily determining the correct order of the en- In this section we briefly describe the ray-casting algo-
try points for the ray. The rendering process follows rithm proposed by Bunyk et al. in [2], that we simply call
Garrity’s method, but when a ray exits the mesh, the al- Bunyk, and the ones proposed by Pina et al. in [13], called
gorithm can easily determine in which cell the ray ME-Ray (Memory Efficient Ray-cast) and EME-Ray (En-
will re-enter the mesh. This approach becomes sim- hanced Memory Efficient Ray-cast).
pler and more efficient than Garrity has proposed, however The data structures used by these algorithms to store the
it keeps some large auxiliary data structures. The re- face data guided their memory usage and performance. The
cent work by Pina et al. [13] proposes two new ray-casting face data are stored in a structure that we call face and cor-
algorithms, ME-Raycast and EME-Raycast, that pre- respond to the information about the geometry and the pa-
sented consistent and significant gains in memory usage rameters of the face. The geometry represents the coordi-
and in the correctness of the final image over Bunyk’s ap- nates of the points that define the face. The parameters are
proach. Their algorithms, however, are slower than Bunyk. costly to compute and represent the constants for the equa-
In terms of the acceleration techniques used in this work, tion of the plane defined by the face, that will be used to
ray coherence has been used earlier for accelerating ray compute the intersection between the ray and the face, and
tracing of traditional surface models [14, 15], and later used the parameters are used to interpolate the scalar value in-
in ray-casting for skiping over empty spaces [8]. side the face. The whole face data is the most consuming
With the advent of programmable graphics hardware, data structure in ray-casting.
several volume rendering algorithms have been developed
taking advantage of this feature to achieve high perfor- 3.1. Bunyk
mance. Graphics hardware based solutions usually provide
real-time performance and high quality images. Weiler et In a preprocessing step, all points and tetrahedra of the
al. [16] present a hardware-based ray-casting algorithm input dataset are read, a list of all vertices and a list of cells
based on the work of Garrity. They find the initial ray are created. In addition, for each input tetrahedron, all four

3. of its faces are kept in a list. Bunyk also maintains, for data structures allow them to deal with all degenerate sit-
each vertex, a list of all faces that use that vertex, called uations that Bunyk is not able to handle.
referredBy list. After reading the dataset, Bunyk deter-
mines which faces belong to the visible side of the scene 4. Visible Face Driven Ray-Casting
boundary, the visible faces. The algorithm projects these
visible faces on the screen, creating for each pixel, a list of Comparing the three ray-casting approaches described in
the intersections, that will be used as the entry point of its the previous section, we can observe that the more faces
ray. After the entry points are stored, the actual ray-casting data are kept in memory, the faster the algorithm computes
begins. For every pixel, the first intersection is its entry point the next intersection of the ray. For this reason, Bunyk is
into the data, and each next intersection is obtained by in- still the fastest software implementation of ray-casting.
specting the intersection between the ray and all other faces Our algorithm, called Visible Face Driven Ray-Casting –
of the current cell. Each pair of intersections are used to VF-Ray, addresses the problem of maintaining information
compute the contribution of the cell for the color and opac- about the faces in memory, minimally degrading the render-
ity of the pixel. The searching for the next intersection in ing performance. The basic idea behind VF-Ray is to keep
Bunyk is quite fast, since it is performed among the other in main memory only the faces of the most recent traver-
three faces of the tetrahedron. However, the list of all faces sals. To do so, we explore ray coherence, using the visi-
and the referredBy lists result in a very high memory ble face information to guide the creation and destruction
consumption. of faces in memory.
3.2. ME-Ray 4.1. Exploring Ray Coherence
ME-Ray starts in a preprocessing step, creating a list of VF-Ray is based on ray coherence, which comes from
all vertices, a list of all cells, for each cell the list of cells the fact that the set of faces intersected by two rays, cast
that share a face with it, and the Use Set list for each ver- from neighboring pixels, will contain almost the same faces.
tex. The Use Set lists substitute the referredBy lists Our algorithm takes advantage of ray coherence by keeping
used in Bunyk, each Use Set contains a list of all cells in- in memory only the faces of nearby rays.
cident on the vertex. In the rendering algorithm, ME-Ray One important issue in exploring ray coherence is the de-
creates a list of visible faces, and determine the entry point termination of the set of rays that are likely to have the same
of each ray, in the same way as Bunyk. The algorithm pro- set of intersected faces. We use the visible faces informa-
ceeds looking for the next intersection, by scanning through tion to do so. The set of rays that may reuse the same list
out the neighbor cells. ME-Ray creates on demand a list of faces are the ones that correspond to the set of pixels un-
of faces, as the face is intersected by the ray. Faces that der the projection of a certain visible face. This set of pixels
are never intersected by any ray will not be created. After will be called here visible set. Figure 1 shows a 2D exam-
each pair of intersection is found, the contribution to the fi- ple of this idea. The rays that are cast through the visible
nal color and opacity is computed. ME-Ray can save about face, presented in the figure, tend to intersect the same in-
40% of the memory used by Bunyk, but is slower, as it has ternal faces. Therefore, the internal faces are created and
to create the faces for the first time during the rendering. stored once, for the first ray intersection. After that, they are
reused until all the pixels of the visible set are computed.
3.3. EME-Ray This process is repeated for all visible faces. To guarantee
correctness, all visible faces are ordered in depth order.
EME-Ray was developed as an optimization of ME-Ray
in terms of memory usage. They have similar behavior. screen
However, EME-Ray does not store the faces in memory,
since it was one of the most memory expensive structures in Visible Face
Projection
ME-Ray. Therefore, EME-Ray has to recalculate the faces Internal Faces
parameters for the computation of the intersections every
time a new face is found. With this optimization, EME-Ray
uses only about 1/4 of the memory used by Bunyk. The sig-
nificant reductions in memory usage comes with an increase
in the execution time, EME-Ray usually doubles Bunyk ex-
ecution time. It is also important to notice that ME-Ray and
Figure 1. Visible face rays coherence.
EME-Ray gains over Bunyk are not only in memory us-
age, but also in the correctness of the final image. Their

4. Our results show that exploiting pixel coherence reduces ated and inserted into computedFaces list. Otherwise,
considerably the memory usage, while keeping the VF- the f acej data are loaded from the list. The insertion and
Ray performance competitive with the fastest previous al- loading operations in the list are done in constant time, since
gorithms because the reuse of face data improves cache per- in the cell we store the index for the position of each face
formance. In fact, for all datasets we tested, approximately in the list. The computation of the intersections is done us-
more than half of the faces are created at most two times for ing the face parameters computed when the face is created.
medium size images. For each intersection, the scalar value is interpolated be-
tween the three vertices of each face, which are used in the
4.2. Data structures illumination integral proposed by Max [10].
The data structures used by VF-Ray are similar to that
of EME-Ray: an array of vertices, an array of cells, and For each visible face vf acei do
the Use Set of each vertex. Besides these structures, VF- Project vf acei ; // Determine visible set
Ray also has a list, called computedFaces that holds the For each pixel p of visible set do
Intersect vf acei ; // Interpolate scalar value
internal faces throughout the model. Note that this list is
Do
cleaned up after the ray-casting computation of all pixels of
Find next internal face f acej ; // Intersection
a visible set. In this way, we don’t need to store all the in- If f acej not in computedF aces then
ternal faces, as done by Bunyk and ME-Ray algorithms. Create f acej and Insert in computedF aces;
Furthermore, just like ME-Ray and EME-Ray, our algo- else
rithm can handle any type of convex cell. For tetrahedral Load f acej from computedF aces;
meshes, the intersections are computed between the line de- Ray integrate; // Optical model
fined by the ray path and the plane defined by the three While exist next face f acej ;
vertices of a triangular face. For hexahedral meshes, where End For
each face is quadrangular, the intersection is found by split- Clear computedF aces; // Clear list
ting these quadrangular faces into two triangular faces, and End For
the same calculation is performed for each one.
Figure 2. The pseudo-code of VF-Ray.
4.3. Handling Degeneracies
The degenerate situations, which can be found during
the ray traversal, are treated in the same way that was per-
formed in Pina et al. [13], by investigating the Use set of
each vertex. So, when a ray hits a vertex or an edge, the al- 5. Results
gorithm can determine the next intersection correctly.
In this section we evaluate the performance and memory
4.4. Algorithm usage of VF-Ray when compared to ME-Ray, EME-Ray
and Bunyk. We also show a comparison between VF-Ray
VF-Ray algorithm starts by reading the input dataset and and a cell projection algorithm, called ZSweep [4]. This last
creating three lists: a list of vertices, a list of cells and the experiment was done in order to put our results in perspec-
Use Set list for each vertex. In the rendering process, VF- tive, with respect to other rendering paradigm. The main
Ray creates a list of all visible faces, that are the faces whose idea of ZSweep algorithm is the sweeping of the data with a
normal makes an angle greater than 90o with the viewing plane parallel to the viewing plane XY , towards the pos-
direction. For each visible face, the algorithm follows the itive z direction. The sweeping process is performed by
steps presented in the pseudo-code of Figure 2. ordering the vertices by their increasing z coordinate val-
Initially, the visible face, vf acei , is projected onto the ues, and then retrieving one by one from this data struc-
screen, generating the visible set of vf acei . For each pixel ture. For each vertex swept by the plane sweep, the algo-
p in the visible set, the algorithm first computes the intersec- rithm projects, onto the screen, all faces that are incident to
tion between the ray cast from p and vf acei . After this first it. During face projection, ZSweep stores a list of intersec-
intersection is found, the ray traversal begins, when the al- tions for each pixel and uses delayed compositing. The pix-
gorithm computes the intersections of the ray with the inter- els lists intersections are composed and the lists are flushed
nal faces. The next face intersected, f acej , is found on an- each time the plane reaches the target-z. The target-z is the
other face of the current cell. If f acej was not created be- maximum z-coordinate among the vertices adjacent to the
fore (this is the first intersection in f acej ), the face is cre- vertex intersected by the sweeping plane.

5. 5.1. Testing Configuration ME-Ray uses from 1.5 to 6 times more memory than VF-
Ray. For EME-Ray, we notice an interesting result. Al-
The VF-Ray algorithm was written in C++ ANSI with- though EME-Ray does not store the faces in memory, for
out using any particular graphics libraries. Our experiments high resolution images, bigger than 1024 × 1024, EME-
were conducted on a Intel Pentium IV 3.6 GHz with 2 GB Ray uses more memory than VF-Ray. This result can be
RAM and 1MB of L2 cache, running Linux Fedora Core 5. explained by the fact that EME-Ray, just like Bunyk and
We have used four different tetrahedral datasets: Blunt ME-Ray, maintains for each pixel a list containing the entry
Fin, SPX Liquid Oxygen Post, and Delta Wing. Table 1 faces for the pixel. The number of lists increases with the in-
shows the number of vertices, faces, boundary faces and crease in image resolution. VF-Ray, on the other hand, does
cells for each dataset. We varied the image sizes, from not require this kind of list, since it computes the entry point
512 × 512 to 8192 × 8192 pixels. The models rendered with of each ray on demand, for each visible face.
VF-Ray are shown in Figures 3 to 6. The images generated When compared to ZSweep, we can observe that except
by VF-Ray, ME-Ray, and EME-Ray are identical, and bet- for SPX with a 512 × 512 image, VF-Ray uses much less
ter than the images generated by Bunyk. There are two rea- memory than ZSweep. For the higher resolution images,
sons for this difference: VF-Ray, ME-Ray, and EME-Ray ZSweep can use 8 or 9 times more memory than VF-Ray.
use double precision variables to compute ray-faces inter- To render SPX with a 8192 × 8192 image, for example,
sections, while Bunyk used float precision, causing artifacts ZSweep spends about 2GB of memory. The ZSweep mem-
in the images, and Bunyk does not handle correctly all de- ory consumption is directly proportional to the final image
generate cases, generating some incorret pixels. resolution, since it creates pixels lists to store the intersec-
tions found for each pixel.
Dataset # Verts # Faces # Tets # Boundary
Blunt 41 K 381 K 187 K 13 K 5.3. Rendering Performance
SPX 149 K 1.6 M 827 K 44 K
Table 3 shows the rendering time (in seconds) of VF-Ray
Post 109 K 1M 513 K 27 K
for Blunt, SPX, Post and Delta, and the relative timing re-
Delta 211 K 2M 1M 41 K
sults of ME-Ray, EME-Ray, Bunyk and ZSweep when com-
Table 1. Dataset sizes. pared to VF-Ray execution. Just like the previous table, the
percentages presented in this table correspond to the ratio
of the other algorithms execution time over VF-Ray execu-
tion time. The time results correspond to the rendering of
5.2. Memory Consumption one frame and do not include preprocessing time.
As we can observe in Table 3, VF-Ray outperforms ME-
Table 2 shows the memory used by VF-Ray to render Ray, EME-Ray and ZSweep for all datasets and all image
one frame (in MBytes) for Blunt, SPX, Post and Delta, and resolutions. For Blunt, Post and Delta, VF-Ray is about 3
the relative memory usage results of ME-Ray, EME-Ray, times faster than EME-Ray and ZSweep, and uses about
Bunyk and ZSweep when compared to VF-Ray usage. The 3/4 of the time used by ME-Ray to render the image. For
percentages presented in this table correspond to the ratio SPX, the gains of VF-Ray against EME-Ray and ZSweep
of the other algorithms memory usage over VF-Ray mem- are smaller, but still high. When compared to Bunyk, for
ory consumption. In other words, we consider VF-Ray re- Blunt and Delta, VF-Ray presents almost the same execu-
sults as 100% and are presenting how much the other algo- tion time, since the difference is only about 10%. For SPX
rithms increase or decrease this result. and Post, VF-Ray is only about 22% slower than Bunyk.
As we can observe in Table 2, in most of the cases, Observing the increase in the rendering time of VF-Ray
VF-Ray uses less memory than the three other algorithms, for different image resolutions, we notice that as the im-
and its memory usage increases linearly with the resolution. age resolution increases, the rendering time increases in the
When compared to Bunyk, VF-Ray uses from 1/3 to 1/6 of same proportion. The gains over the other algorithms are al-
the memory spent by Bunyk. These gains are impressive. most the same for different resolutions.
For Blunt, for example, VF-Ray renders a 4096 × 4096 im-
age using the same amount of memory that Bunyk uses to 6. Discussion
render a 512 × 512 image. For the largest dataset, Delta,
Bunyk uses more memory to render a 512 × 512 image than VF-Ray had obtained consistent and significant gains in
VF-Ray uses to render a 8192 × 8192 image. memory usage over Bunyk, providing almost the same per-
When compared with the two memory-aware ray-casting formance (the difference on their execution time is only
algorithms, ME-Ray and EME-Ray, we can observe that about 16%). As the image resolution increases, our gains

6. Dataset Resolution Memory (MB) ME-Ray EME-Ray Bunyk ZSweep
512 × 512 14 404% 166% 528% 208%
1024 × 1024 30 240% 129% 297% 177%
Blunt 2048 × 2048 39 333% 251% 377% 369%
4096 × 4096 75 495% 451% 517% 689%
8192 × 8192 219 610% 655% 617% 917%
512 × 512 96 215% 74% 299% 87%
1024 × 1024 99 234% 88% 305% 107%
SPX 2048 × 2148 108 271% 141% 337% 188%
4096 × 4096 144 383% 284% 428% 407%
8192 × 8192 288 533% 498% 569% 711%
512 × 512 63 165% 74% 290% 97%
1024 × 1024 66 203% 95% 301% 129%
Post 2048 × 2048 74 281% 172% 353% 238%
4096 × 4096 110 424% 349% 470% 499%
8192 × 8192 256 601% 560% 603% 800%
512 × 512 116 167% 74% 303% 97%
1024 × 1024 119 200% 84% 308% 116%
Delta 2048 × 2048 129 246% 124% 330% 177%
4096 × 4096 165 346% 247% 403% 375%
8192 × 8192 307 500% 467% 534% 704%
Table 2. Memory usage results for VF-Ray compared to ME-Ray, EME-Ray, Bunyk and ZSweep.
in memory usage over Bunyk are even more pronounced, execution time come from the good cache performance of
but the difference in the rendering time remains the same. VF-Ray. We have made some experiments to evaluate the
For 8192 × 8192 images, Bunyk uses about 6 times more cache misses of VF-Ray and ME-Ray and obtained for VF-
memory than VF-Ray with a performance difference of Ray a cache miss ratio very low, near 0%, while the cache
only about 18%. In our experiments, however, we are only miss ratio of ME-Ray is about 1.9% for modest size im-
comparing executions where the whole dataset fits in main ages. This result was obtained because our algorithm reuse
memory for all algorithms. Our experimental platform has the faces for the rays in the same visible face. Given the fact
2 GB of main memory that can store all the data structures that rays are shot one after the other, the closer the neigh-
used by our workload. As the memory usage increases, the boring rays are to each other, the higher the probability is
rendering will need to use the virtual memory mechanisms that they reuse the cached data.
of the operating system, which would have great influence The comparison of VF-Ray with the cell projection ap-
on the overall execution time. proach, implemented by ZSweep, showed that VF-Ray is
The small increase in the execution time, when compared faster and spends less memory than ZSweep. The gains
to Bunyk, comes from the fact that Bunyk computes all in- are huge. For the Delta dataset, ZSweep uses 7 times more
ternal faces in the preprocessing step, before the ray-casting memory, running about 3.5 times slower than VF-Ray. The
loop, while VF-Ray does it on demand, as the face is inter- pixel list maintained by the ZSweep approach increases
sected for the first time. In addition, in VF-Ray, if a face is while a target-Z was not reached. As this list grows, the
intersected by rays of two different visible faces, it has its insertion is more expensive, since the pixel list must be or-
data computed twice. dered. On the other hand, the data structures on the VF-Ray
The comparison of VF-Ray with ME-Ray and EME- algorithm does not depend on ordering. The difference in
Ray provided interesting results. ME-Ray and EME-Ray memory usage of ZSweep is due to the increase in these
are memory-aware algorithms designed to reduce the great pixels lists. For ZSweep, as the image size increases, each
memory usage of Bunyk. VF-Ray was designed to reduce face projected will insert intersection units into more pixel
ME-Ray memory usage, close to the usage of EME-Ray, but lists. If the target-Z is not appropriate, the pixels lists will
running much faster than EME-Ray. Our results showed, increase considerably.
however, that for high resolution images, our algorithm is In terms of the dataset characteristics, the performance
not only faster than ME-Ray but uses less memory than of VF-Ray increases as the number of pixels in each vis-
EME-Ray. The gains of VF-Ray over EME-Ray in mem- ible face increases. Datasets with a small number of faces
ory usage come from the elimination of the lists of entry have a greater amount of pixels per visible face. Therefore,
faces for each pixel. The gains of VF-Ray over ME-Ray in datasets, like Blunt, that has a small number of faces, tend

7. Dataset Resolution Time (sec) ME-Ray EME-Ray Bunyk ZSweep
512 × 512 1.9 139% 296% 105% 253%
1024 × 1024 7.0 143% 317% 94% 431%
Blunt 2048 × 2048 27.2 146% 337% 88% 481%
4096 × 4096 107.4 147% 328% 88% 516%
8192 × 8192 426.7 154% 341% 91% 382%
512 × 512 4.1 111% 161% 76% 287%
1024 × 1024 13.0 103% 203% 81% 202%
SPX 2048 × 2048 46.6 103% 225% 83% 219%
4096 × 4096 177.1 105% 234% 82% 226%
8192 × 8192 696.3 105% 238% 81% 260%
512 × 512 5.0 138% 251% 80% 201%
1024 × 1024 19.5 136% 254% 76% 167%
Post 2048 × 2048 78.6 133% 258% 74% 194%
4096 × 4096 309.9 135% 257% 74% 227%
8192 × 8192 1246.3 135% 263% 72% 246%
512 × 512 3.1 130% 242% 91% 465%
1024 × 1024 11.2 122% 270% 88% 271%
Delta 2048 × 2048 41.9 121% 280% 87% 312%
4096 × 4096 162.7 120% 290% 84% 341%
8192 × 8192 640.3 123% 290% 83% 369%
Table 3. Time results for VF-Ray compared to ME-Ray, EME-Ray, Bunyk and ZSweep.
to present better performance results for VF-Ray.
Finally, it is important to keep in mind that the reduc-
tions in memory usage will allow the implementation of the
algorithm in the graphics hardware. Furthermore, these re-
ductions will also allow the use of double precision in the
parameters to calculate the intersection between a ray and a
face, raising the chance to obtain more precise images.
Figure 4. SPX image generated by VF-Ray.
dral or hexahedral datasets, including non-convex and dis-
connected meshes even with holes.
Our idea was to explore ray coherence in order to reduce
the memory requirements for the storing of faces data while
achieving a good cache performance. We compared VF-Ray
Figure 3. Blunt image generated by VF-Ray.
memory usage and performance with previous ray-casting
approaches, and with a cell projection algorithm. When
compared to the fastest and memory expensive approach, by
Bunyk, VF-Ray obtained comparable performance, spend-
7. Conclusions ing only from 1/3 to 1/6 of the memory. When compared
to the memory-aware Pina et al algorithms, VF-Ray ob-
In this work, we have proposed a novel memory-aware tained consistent and significant gains in memory usage and
and efficient software ray-casting algorithm for volume ren- in execution time. When compared to the cell projection ap-
dering, called VF-Ray. The algorithm is suitable for tetrahe- proach, VF-Ray also reduced significantly the memory con-

8. [2] P. Bunyk, A. Kaufman, and C. Silva. Simple, fast, and ro-
bust ray casting of irregular grids. Advances in Volume Visu-
alization, ACM SIGGRAPH, 1(24), July 1998.
[3] R. Espinha and W. Celes. High-quality hardware-based ray-
casting volume rendering using partial pre-integration. In
SIBGRAPI ’05: Proc. of the XVIII Brazilian Symposium on
Computer Graphics and Image Processing, October 2005.
[4] R. Farias, J. Mitchell, and C. Silva. Zsweep: An efficient and
exact projection algorithm for unstructured volume render-
ing. In 2000 Volume Visualization Symposium, pages 91 –
99, October 2000.
[5] M. P. Garrity. Raytracing irregular volume data. In VVS ’90:
Proceedings of the 1990 workshop on Volume visualization,
pages 35–40. ACM Press, 1990.
Figure 5. Post image generated by VF-Ray. [6] L. Hong and A. Kaufman. Accelerated ray-casting for curvi-
linear volumes. In VIS ’98: Proceedings of the conference
on Visualization ’98, pages 247–253, 1998.
[7] K. Koyamada. Fast ray-casting for irregular volumes. In
ISHPC ’00: Proc. of the Third International Symposium on
High Performance Computing, pages 557–572, 2000.
[8] S. Lakare and A. Kaufman. Light weight space leaping us-
ing ray coherence. In Proc. IEEE Visualization 2004, pages
19–26, 2004.
[9] R. Marroquim, A. Maximo, R. Farias, and C. Esperanca.
Gpu-based cell projection for interactive volume rendering.
In SIBGRAPI ’06: Proc. of the Brazilian Symposium on
Computer Graphics and Image Processing, October 2006.
[10] N. Max. Optical models for direct volume rendering.
IEEE Transactions on Visualization and Computer Graph-
ics, 1(2):99–108, 1995.
[11] M. Meibner, J. Huang, D. Bartz, K. Mueller, and R. Craw-
Figure 6. Delta image generated by VF-Ray. fis. A practical evaluation of popular volume rendering algo-
rithms. In VVS ’00: Proceedings of the 2000 IEEE sympo-
sium on Volume visualization, pages 81–90, 2000.
sumption and the execution time. [12] A. Neubauer, L. Mroz, H. Hauser, and R. Wegenkittl. Cell-
The reductions in memory usage makes VF-Ray suitable based first-hit ray casting. In VISSYM ’02: Proceedings of the
for using double precision in the parameters to calculate the symposium on Data Visualisation 2002, pages 77–ff, 2002.
intersection between a ray and a face. Using double preci- [13] A. Pina, C. Bentes, and R. Farias. Memory efficient and ro-
sion avoids the artifacts that appear when float precision is bust software implementation of the raycast algorithm. In
used. We conclude that VF-Ray is an efficient ray-casting WSCG’07: The 15th Int. Conf. in Central Europe on Com-
puter Graphics, Visualization and Computer Vision, 2007.
algorithm that allows the in-core rendering of big datasets.
[14] A. Reshetov, A. Soupikov, and J. Hurley. Multi-level ray
As future work, we intend to optimize VF-Ray by keeping
tracing algorithm. In SIGGRAPH ’05: ACM SIGGRAPH
face data for neighboring visible faces, and to parallelize
2005 Papers, pages 1176–1185, 2005.
VF-Ray in order to run on a cluster of PCs [15] I. Wald, P. Slusallek, C. Benthin, and M. Wagner. nterac-
tive rendering with coherent ray tracing. In Eurographics 01
8. Acknowledgments Proceedings, pages 153–164, 2001.
[16] M. Weiler, M. Kraus, M. Merz, and T. Ertl. Hardware-based
We acknowledge the grant of the first and second authors ray casting for tetrahedral meshes. In Proceedings of the 14th
provided by Brazilian agencies CAPES and CNPq. IEEE conference on Visualization ’03, pages 333–340, 2003.
[17] M. Weiler, P. Mall´ n, M. Kraus, and T. Ertl. Texture-encoded
o
tetrahedral strips. In 2004 IEEE Symposium on Volume Visu-
References alization and Graphics (VolVis 2004), pages 71–78, October
2004.
[1] F. Bernardon, C. Pagot, J. Comba, and C. Silva. Gpu-based
tiled ray casting using depth peeling. Journal of Graphics
Tools, To appear.